compute so slowly

[junedgar]

I ran the video_tracking.py with GPU. But it worked so slowly. The log's output is below:

Time required: 860.9sec
6235 vertices and 1 classes and 6536778 edges
starting objective: -4550.940549
....pair updates: 100.195555
....new sets: 386.270724
....updating class labels: 0.000000
..new objective: -5037.406828
....pair updates: 32.952381
....new sets: 0.000000
....updating class labels: 0.000000
..new objective: -5070.359209
....pair updates: 6.934516
....new sets: 0.000000
....updating class labels: 0.000000
..new objective: -5077.293725
....pair updates: 0.159989
....new sets: 0.000000
....updating class labels: 0.000000
..new objective: -5077.453715
....pair updates: 0.232185
....new sets: 0.000000
....updating class labels: 0.000000
..new objective: -5077.685900
Objective: -5077.685900
num_people:  1839
Object 0 tracked at [666,275, 735,546]
Object 1 tracked at [632,25, 705,349]
Object 2 tracked at [124,451, 232,683]
Object 3 tracked at [568,285, 671,514]
Object 4 tracked at [389,315, 443,628]
is_new_person!
is_new_person!
People(this frame): 7
People(cumulative): 7
Frame: 1/72
Time required: 1674.7sec
6186 vertices and 1 classes and 6411429 edges
starting objective: -4521.743644
....pair updates: 72.095567
....new sets: 417.133020
....updating class labels: 0.000000
..new objective: -5010.972232

....pair updates: 27.561003
....new sets: 1.150275
....updating class labels: 0.000000
..new objective: -5039.683510

Did i run the demo with wrong way? I followed the guidance "python3 video_tracking.py -f 'test_video_01f.mov'".

[chituma110]

@junedgar I have meet the same problem,but I run the code with CPU the time required is 577.5 sec. How did you run the code with GPU?

[junedgar]

@chituma110 I had not explored the code completely. But i found the code initialized the Session without any limitations for GPU, so i thought that the code may use the GPU directly。

[gg-z]

I've met the same problem, I solve this by cloning the original repo: pose_tensorflow and copying the tracking-related code to repo:pose_tensorflow. You can give it a try. it works for me. I guess that problem comes from some configuration of pose_tensorflow used in this project.